/StyleGAN-Tensorflow

Simple & Intuitive Tensorflow implementation of StyleGAN (CVPR 2019 Oral)

Primary LanguagePythonMIT LicenseMIT

StyleGAN-Tensorflow

Simple & Intuitive Tensorflow implementation of "A Style-Based Generator Architecture for Generative Adversarial Networks" (CVPR 2019 Oral)

Other implementation

Usage

├── dataset
   └── YOUR_DATASET_NAME
       ├── 000001.jpg 
       ├── 000002.png
       └── ...

Train

> python main.py --dataset FFHQ --img_size 1024 --gpu_num 4 --progressive True --phase train

Test

> python main.py --dataset FFHQ --img_size 1024 --progressive True --batch_size 16 --phase test

Draw

Figure02 uncurated

python main.py --dataset FFHQ --img_size 1024 --progressive True --phase draw --draw uncurated

Figure03 style mixing

python main.py --dataset FFHQ --img_size 1024 --progressive True --phase draw --draw style_mix

Figure08 truncation trick

python main.py --dataset FFHQ --img_size 1024 --progressive True --phase draw --draw truncation_trick

Architecture

Our Results (1024x1024)

  • Training time: 2 days 14 hours with V100 * 4
  • max_iteration = 900
    • Official code = 2500

Uncurated

Style mixing

Truncation trick

Generator loss graph

Discriminator loss graph

Author

Junho Kim